Where ChatGPT helps with SEO, where it does nothing, and where it costs you
A prompt that sorts your own SEO tasks into three buckets: the ones a model does well, the ones where it only saves typing, and the ones where it hurts.
- Works in
- ChatGPT, Claude, Gemini
- You need
- A list of the SEO tasks you currently do · Which of them you already hand to a model · One sentence on what you sell
- Written for
- how to use chatgpt for seo
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
Most answers to this question are written by people selling something, which is why so many people add “reddit” to the query. The honest version has three parts: a short list of things a model genuinely does better, a longer list where it saves typing, and a list where it quietly costs you. The prompt below sorts your own tasks into those three, and the rules stop it being generous with itself.
What it is actually good at
Reasoning about text you paste. That is the whole of bucket one, and it is a real capability rather than a consolation.
Give a model 300 keywords and it groups them by intent better than a spreadsheet ever will. Give it a page and it will tell you the answer arrives in paragraph four. Give it a crawl export or a robots.txt somebody inherited and it will read the machine output and say what it means. Give it eight competitor titles and it will name the pattern all eight share. Every one of those has its evidence in the prompt, and none of them requires the model to know anything about the live web.
What it cannot do, at all
Everything that needs data. A model has no search volume, no rank data, no backlink profile, no access to your analytics and no sight of a live results page.
That list is not a prompt engineering problem. There is no wording that gives a model a number it does not have, and the failure mode is specific: asked to prioritise, it produces plausible figures rather than refusing. Content plans get built on volumes no tool ever returned. The same goes for any claim about the current web, which it cannot verify and will assert anyway.
Where it costs you without announcing it
Unedited prose. Model output is fluent, well organised, and made almost entirely of general statements, which is exactly what a page needs least.
There is no reliable public benchmark for detecting generated text and we are not going to invent one. The practical problem is simpler and it has nothing to do with detection: a page with no specifics, no numbers you measured and no position anybody could argue with does not get read, quoted or linked, whoever wrote it. The specifics are yours to supply. That is the part the model cannot do for you, and it is most of the value.
Start with what no prompt can see for you: run the indexability checker on a page you care about, then work through the analysis prompts for the jobs that sit inside what a model can actually reason about.
You are auditing a language model, and the language model is you. I am going
to list the SEO tasks I do. Sort them by whether handing them to a model is
worth doing, and be hard on yourself.
What you are: a system that reads text I paste, reasons about structure and
intent, and writes fluently. What you are not: anything with access to search
volume, live search results, rankings, backlinks, my analytics, my crawl data
or my competitors. You cannot verify a single claim about the live web and
you produce confident output regardless of whether you have grounds for it.
That is the failure you are auditing for.
Sort every task I list into exactly one of three buckets.
BUCKET ONE, genuinely better with a model. Only tasks where all the evidence
needed is text I can paste, and where the work is reasoning about that text
or shaping it. For each: what specifically you do better than a person under
time pressure, and the one instruction that has to be in the prompt for it to
hold.
BUCKET TWO, saves typing and nothing else. Tasks where you produce a first
draft of similar quality to what I would write, faster. Say plainly that the
gain is speed, not quality, and name what I still have to do afterwards.
BUCKET THREE, actively costs me something. Tasks where you will produce
output that looks finished and is wrong, unverifiable or generic. For each:
the specific failure, what it looks like when it happens, and what I should
use instead. Name a real category of tool where one exists.
Rules you must follow:
1. Any task that needs search volume, rank data, backlink data, a live search
results page, crawl output or analytics goes in bucket three. There is no
prompt wording that moves it out. Say so rather than suggesting a better
prompt.
2. Any task whose output I cannot check in under five minutes goes in bucket
three, even when you would probably get it right. Unverifiable output at
scale is the expensive failure, not the occasional wrong answer.
3. Do not soften bucket three. No "with careful prompting" and no "as a
starting point". If it belongs there, say why in plain words.
4. Bucket one must be shorter than bucket three, or you have been generous
with yourself. If your honest reading makes it longer, say explicitly why
this list is unusual.
5. Do not invent statistics, studies, percentages or accuracy figures about
AI writing, AI detection or model performance. If you reach for a number,
you are guessing. Leave it out.
6. Finish with the tasks I did not list that belong in bucket one, and with
any task I said I already hand to a model that you have placed in bucket
three. Put that second list first if it is not empty.
Format: three headed buckets, each task as one short block. No table.
What I sell, and to whom: [ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]
Tasks I currently do: [LIST YOUR SEO TASKS, ONE PER LINE]
Tasks I already hand to a model: [LIST THEM, OR WRITE "none"]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]- Changes which tasks matter enough to argue about. A local business with eight pages and an ecommerce site with nine thousand have almost nothing in common in this audit, and without the sentence you get the generic answer written for neither.
[LIST YOUR SEO TASKS, ONE PER LINE]- Everything you actually do in a month, written the way you would say it out loud. "Write meta descriptions for new products", "work out why traffic dropped", "check the new pages are indexed". Vague entries like "content" produce vague verdicts, so split anything that is really three jobs.
[LIST THEM, OR WRITE "none"]- The tasks you have already automated with a model. This is the input that makes the audit uncomfortable and useful, because rule 6 forces the model to open with any of them it has just placed in bucket three.
How to run it
- 01Write the task list before you read anything about AI
List what you genuinely do rather than what an SEO workflow diagram says you should. The audit is only as good as the list, and a borrowed list produces verdicts about a job that is not yours. Fifteen to twenty five lines is the useful range.
- 02Be honest in the second list
Write down every task you have quietly handed to a model, including the ones you would not mention in a meeting. Rule 6 makes the model surface any of them it considers harmful before anything else, and that section is the entire reason to run this.
- 03Run it and start at bucket three
Read the harm bucket first and read it slowly. It is the part a vendor page will not write for you, and it is where the tasks live that produce finished looking output nobody ever checks. Bucket one will still be there afterwards.
- 04Test one bucket one claim immediately
Take a single task the model put in bucket one and run it now on real input of yours. Compare the output to what you would have produced. One test tells you more about your own workflow than the whole audit does, because model quality varies by task in ways no general answer captures.
- 05Replace bucket three with the tool category it named
Every bucket three entry should point at a real category of tool: a rank tracker, a crawler, an analytics export, Search Console. If a task is in that bucket and you have no tool for it, you have found a genuine gap, and no prompt is going to fill it.
- 06Settle the things a model can never see for you
The tasks that need live data start with the plumbing. Run your URL through the free indexability check to confirm the page is even allowed in the index. No model can read that from anything you paste, and it overrides every content judgement it makes.
Questions people ask
How should I use ChatGPT for SEO?
For reasoning about text you already have, not for facts about the web. Clustering a keyword export by intent, turning a page into a clean outline, drafting titles against a real character limit and reading a crawl export are all jobs where the evidence sits in the prompt. Anything needing volume, rankings or links needs a tool that measures them.
What can ChatGPT not do for SEO?
It cannot see a live search results page, your analytics, your backlink profile or any current ranking, and it has no search volume data. It also cannot verify a claim about the live web. Asked for any of those it produces figures that look exactly like real ones, which is worse than refusing, because nobody checks a number that arrives with confidence.
Is AI generated content penalised by Google?
Google says it judges content by quality and usefulness rather than by how it was produced, and low value pages made at scale are the stated target regardless of method. In practice the risk is not the tool, it is that unedited model output is generic, unsourced and says nothing a hundred other pages do not, which is what actually fails.
Can I use ChatGPT to write blog posts?
You can draft with it, and the draft is the cheap part. Model prose defaults to fluent generality: no specifics, no numbers you measured, no opinion anybody could disagree with. That is fixable only by supplying the specifics yourself, which is the work you were hoping to avoid. Used as an editor rather than an author it earns its place.
Why does the prompt make the model criticise itself?
Because asked neutrally it is agreeable, and an agreeable answer to this question is an advert. Forcing three buckets, requiring the harm bucket to be at least as long as the useful one and banning softening language produces a list you can act on rather than a list of encouragements.
Unsubscribe in one click. We never pass your address on.